fix(provider): support custom provider image modalities

[saiqulhaq-hh]

Issue for this PR

Closes #47480

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Custom-provider models can reach the message filter with capabilities.input in the modalities array form, such as ["text", "image"]. The filter only checked the object form (input.image), so it replaced valid image parts with a "model does not support image input" error before calling the provider.

This change accepts both capability representations at that boundary: array capabilities use includes(modality), while registry-derived object capabilities keep the existing keyed lookup. A regression test covers an image-capable custom-provider model using the array form.

How did you verify your code works?

• bun test test/provider/transform.test.ts from packages/opencode (586 passed)
• bun typecheck from packages/opencode
• bun run test from packages/opencode (3612 passed, 22 skipped, 1 todo)
• Prettier and oxlint on the two changed files

I also ran the patched CLI with bun dev . and used the Read tool on a local image with this sanitized custom-provider configuration:

{
  "provider": {
    "litellm": {
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "baseURL": "https://<litellm-host>/v1",
        "apiKey": "{env:LITELLM_API_KEY}"
      },
      "models": {
        "minimax/MiniMax-M3": {
          "name": "MiniMax M3",
          "attachment": true,
          "modalities": {
            "input": ["text", "image"]
          }
        }
      }
    }
  }
}

The model successfully parsed text from the image instead of receiving the unsupported-image error.

Screenshots / recordings

Local test result: bun dev .

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

If you do not follow this template your PR will be automatically rejected.

[github-actions]

The following comment was made by an LLM, it may be inaccurate: